Open d1 Open-weight multimodal decision models from Liquid AI. d1-3B is a 3B vision-language model that returns calibrated yes/no, routing, or ranking answers in a single forward pass instead of generating tokens. d1-omni-600M is a 600M experimental model that also accepts audio. Both run locally on NVIDIA DGX, RTX, and Jetson hardware with llama.cpp support, targeting agent guardrails, content moderation, and on-device intent classification.
Visit productFeatured media
Loading media…
View on XCosigns Received
“@liquidai congrats on the open weights drop 🔥 already making quants for this :)”
“open d1 decision models are here! • small enough to run them on my macbook • multimodal for fun things like video input here i'm using d1-3B to rate how cute my outfit is. don't follow me for more fashion advice 💜 see more fun demos on our hugging face space:”
“Tiny decision models you can run everywhere: > d1-3B has impeccable performance in text and vision > d1-omni-600M supports text, vision, and audio! Try our @huggingface space with interactive games today”
“I ported @LiquidAI’s d1-omni-600M to MLX for Apple Silicon. Text, images & audio → structured decisions on your Mac. 1.17 GB FP16. ~80 ms for 3 text questions on M3 (warm). 79/79 checked decisions matched the original.”
“Run Liquid AI's d1 decision models locally ⚖️ We dropped GGUF quants of d1-3B and d1-omni-600M on Hugging Face, from 256 MB to 5.4 GB The sweet spot is our 1.66 GB d1-3B AD-Q4_K_M and it's top-1% is 97.1% compared to BF16 GGUFs: https://huggingface.co/AtomicChat/d1-3B-GGUF”
“d1-omni-600M, now running on Apple silicon with Core ML ⚡ We pointed it at a firehose of 5,000 real comments from Civil Comments with GPU + ANE. moderating 224 comments/s on an M5 Pro. 95.4% agreement with human raters, same as the base PyTorch too. 🤗 Model: https://t.co/oxwfoHjuvr 🛠️ SDK:”
“running d1-3B on a single RTX 3060. 12GB VRAM. (config below) 88.4% on BoolQ (full val, 3,270 questions), Liquid's card says 86.7. 44.5% on MMLU-Pro (all 12,032, random is 10%). ~43 ms p50 per decision, 21-23 decisions/s, 5 GB VRAM. GPQA lands at 29.8%, close to random, which fits Liquid's own Decision Index showing knowledge as its weakest area. added a decision harness to BenchKit for this. calibration next! config (llama.cpp PR #30110, merged): llama-server -m d1-3B-Q8_0.gguf --mmproj mmproj-d1-3B-F16.gguf -ngl 99 -c 32768 -fa on -np 1 -b 4096 -ub 4096”